The AI PHI Affinity Group’s July talk hosts Ehsan Adeli, PhD — Assistant Professor of Psychiatry & Behavioral Sciences at Stanford University, with courtesy appointments in Biomedical Data Science and Computer Science. Adeli directs Stanford’s Translational AI Lab (STAI) and co-directs Stanford AI for Mental Health (AI4MH). His talk, “MIRAGE: The Illusion of Visual Understanding,” draws on his lab’s recent work documenting a class of failure mode that should make anyone deploying vision-language models in medicine pause.

The MIRAGE work shows that frontier multimodal models will confidently produce detailed image descriptions — including pathology-laden clinical findings — for images that were never actually provided to the model. The same models can score near the top of multimodal benchmarks while having no access to the underlying images at all. When the prompt explicitly invites a guess in the absence of an image, performance drops sharply, suggesting these systems default to confident fabrication rather than conservative refusal whenever the framing leaves room for it. Adeli and co-authors propose B-Clean as a vision-grounded evaluation protocol — a finding with direct implications for any team measuring multimodal AI performance on clinical imaging.

The talk is Friday, July 10, 2026 at 9:00 AM Hawaiʻi time, on Zoom. (July 10 is the second Friday of the month; Independence Day is observed on Friday, July 3.)

Register for this Affinity Group meeting →

Register once and Zoom will email you the join link plus reminders for this and any future Affinity Group meeting. Recordings are posted to the AI PHI YouTube channel ↗. To suggest a future speaker, email aiphi@shepherdresearchlab.org.